If you printed and stacked the pages @reductoai has processed this year, the stack would reach space multiple times over We just wrapped an insane Q2 but some highlights that stand out: - 🖥️ We made a lot of progress on core model performance and grew the product surface. Reducto is now faster and more accurate across the board - 📈 Our GTM team is crushing it, and the team delivered 350% to quota capacity this quarter - 🖥️ (pt 2) Our new products like Deep Extract are setting a new standard for the industry. In @micro1_ai's benchmark there was literally no solution that came close. - 💰We broke our record for largest new contract twice - 🏢 We're now 50% over intended capacity for our office, and are moving to a beautiful new space soon --- The day to day of startup life is always chaotic with plenty of ups and downs, but every time I pause to think about the bigger picture of our journey it's genuinely hard to believe. Revenue growth is still around 8x YoY, and we're just now ramping our capacity. Our sales team in Q3 will be 4x the size of what it was in Q2. Even more importantly, we secured a massive cluster of GPUs and there's a lot that our product research teams are working on that I think will be a step change for our customers. As always, could not be more grateful for our customers and for the team that makes all of that possible. More from us soon 🔮

There is a lot to see at AIE this year, but I feel like this is inherently unique

Thanks @aiDotEngineer for the opportunity to speak! Appreciate everyone that attended and listened to me yap about agents for a bit :)

agi is here

Today we're publishing LongExtractBench, a benchmark commissioned by @reductoai and independently validated by micro1. We evaluated seven production document extraction systems across the same 225 complex enterprise documents. The benchmark was intentionally difficult: documents averaged 358 pages and contained roughly 88,700 ground-truth fields each. Every system was evaluated using the configuration documented in the benchmark methodology. Key findings: • Reducto Deep Extract was the only system to successfully complete all 225 documents. • Direct frontier LLM baselines achieved substantially lower completion rates on long, complex documents. • In this benchmark, dedicated extraction platforms achieved higher completion rates than the direct frontier LLM baselines. • Recall was the clearest differentiator. Precision remained high across systems, but recall ranged from 33.8% to 99.6%, highlighting which systems consistently captured the information contained in long, complex documents. The full report includes the benchmark methodology, limitations, and reproducibility resources. Check out the report and results in the comments below.

How can we log, audit, and improve deployed AI models and agents for medicine? 🏥 AI is rapidly entering the clinic. Although models are already deployed across millions of patients, there is no consistent approach to monitor model use, assess performance, detect failure modes, or measure clinical impact. Model cards and data sheets describe a model before deployment. After deployment, there is no standard to record how, when, by whom, and for whom health AI models are used. We need a “black box” flight recorder for medical AI. ✈️ Co-supervised by @zakkohane and @marinkazitnik, I led a team across 53 institutions and 11 countries to develop MedLog, a protocol for event-level logging of medical AI. Each time an AI model makes a prediction, MedLog creates a record with nine core fields: header, model, user, target, inputs, artifacts, outputs, outcomes, and feedback. In our updated pre-print, we now report four real-world pilot implementations of MedLog across the US, Switzerland, and Vietnam. Predictive and generative AI, with 3M logged model invocations across up to 791K patients. 🇨🇭ICU deterioration monitoring in Switzerland: 223,840 records, 114 days, 212 patients 🇻🇳Wearable tetanus progression monitoring in Vietnam: 3,406 records, 289 days, 15 patients 🇺🇸LLM-based sepsis quality reporting in San Diego: 3,766 records, 89 days, 60 patients 🇺🇸Patient non-attendance prediction in New York: 2,914,264 records, 244 days, 791,319 patients A thread on what MedLog revealed and how to get involved 🧵👇 1/n

make your agents smarter 🔜

New reading material for my coffee table just dropped @aiDotEngineer

As much as other "agent managers" try to push parallelized agents (and even Cursor itself with the Agent window), the flexibility of the Cursor Editor has been so nice to stay chaotically organized

We made JARVIS for your computer. An operating system you control with your voice. The mouse and keyboard had a good run.

good events are an art form

imagine telling your customers there's a small chance you'll randomly decide they're using your product wrong and you won't tell them but will secretly silently sabotage their work

autoresearch

southwest should be paying ME to show up at 6am at LAX

a good sunday brunch

In building any AI tooling that deals with ambiguity like this facing users I’ve usually had to put guardrails that told the models to claim uncertainty rather than creating terrible guesses

highly recommend the cardamom cappuccino 👀

The DoD dumped decades worth of UFO reports. We parsed and open sourced all of the documents Have fun 👽